; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 -mattr=+real-true16 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GFX12-TRUE16 %s
; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 -mattr=-real-true16 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GFX12-FAKE16 %s

declare bfloat @llvm.sqrt.bf16(bfloat %a)
declare <2 x bfloat> @llvm.sqrt.v2bf16(<2 x bfloat> %a)

define amdgpu_kernel void @sqrt_bf16(ptr addrspace(1) %r, ptr addrspace(1) %a) {
; GFX12-TRUE16-LABEL: sqrt_bf16:
; GFX12-TRUE16:       ; %bb.0: ; %entry
; GFX12-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
; GFX12-TRUE16-NEXT:    s_mov_b32 s6, -1
; GFX12-TRUE16-NEXT:    s_mov_b32 s7, 0x31016000
; GFX12-TRUE16-NEXT:    s_mov_b32 s10, s6
; GFX12-TRUE16-NEXT:    s_mov_b32 s11, s7
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    s_mov_b32 s8, s2
; GFX12-TRUE16-NEXT:    s_mov_b32 s9, s3
; GFX12-TRUE16-NEXT:    s_mov_b32 s4, s0
; GFX12-TRUE16-NEXT:    buffer_load_u16 v0, off, s[8:11], null
; GFX12-TRUE16-NEXT:    s_mov_b32 s5, s1
; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT:    v_sqrt_bf16_e32 v0.l, v0.l
; GFX12-TRUE16-NEXT:    buffer_store_b16 v0, off, s[4:7], null
; GFX12-TRUE16-NEXT:    s_endpgm
;
; GFX12-FAKE16-LABEL: sqrt_bf16:
; GFX12-FAKE16:       ; %bb.0: ; %entry
; GFX12-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
; GFX12-FAKE16-NEXT:    s_mov_b32 s6, -1
; GFX12-FAKE16-NEXT:    s_mov_b32 s7, 0x31016000
; GFX12-FAKE16-NEXT:    s_mov_b32 s10, s6
; GFX12-FAKE16-NEXT:    s_mov_b32 s11, s7
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    s_mov_b32 s8, s2
; GFX12-FAKE16-NEXT:    s_mov_b32 s9, s3
; GFX12-FAKE16-NEXT:    s_mov_b32 s4, s0
; GFX12-FAKE16-NEXT:    buffer_load_u16 v0, off, s[8:11], null
; GFX12-FAKE16-NEXT:    s_mov_b32 s5, s1
; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x0
; GFX12-FAKE16-NEXT:    v_sqrt_bf16_e32 v0, v0
; GFX12-FAKE16-NEXT:    buffer_store_b16 v0, off, s[4:7], null
; GFX12-FAKE16-NEXT:    s_endpgm
entry:
  %a.val = load bfloat, ptr addrspace(1) %a
  %r.val = call bfloat @llvm.sqrt.bf16(bfloat %a.val)
  store bfloat %r.val, ptr addrspace(1) %r
  ret void
}

define amdgpu_kernel void @sqrt_v2bf16(ptr addrspace(1) %r, ptr addrspace(1) %a) {
; GFX12-TRUE16-LABEL: sqrt_v2bf16:
; GFX12-TRUE16:       ; %bb.0: ; %entry
; GFX12-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
; GFX12-TRUE16-NEXT:    s_mov_b32 s6, -1
; GFX12-TRUE16-NEXT:    s_mov_b32 s7, 0x31016000
; GFX12-TRUE16-NEXT:    s_mov_b32 s10, s6
; GFX12-TRUE16-NEXT:    s_mov_b32 s11, s7
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    s_mov_b32 s8, s2
; GFX12-TRUE16-NEXT:    s_mov_b32 s9, s3
; GFX12-TRUE16-NEXT:    s_mov_b32 s4, s0
; GFX12-TRUE16-NEXT:    buffer_load_b32 v0, off, s[8:11], null
; GFX12-TRUE16-NEXT:    s_mov_b32 s5, s1
; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT:    v_sqrt_bf16_e32 v1.l, v0.l
; GFX12-TRUE16-NEXT:    v_nop
; GFX12-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(TRANS32_DEP_2)
; GFX12-TRUE16-NEXT:    v_sqrt_bf16_e32 v0.l, v0.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
; GFX12-TRUE16-NEXT:    buffer_store_b32 v0, off, s[4:7], null
; GFX12-TRUE16-NEXT:    s_endpgm
;
; GFX12-FAKE16-LABEL: sqrt_v2bf16:
; GFX12-FAKE16:       ; %bb.0: ; %entry
; GFX12-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
; GFX12-FAKE16-NEXT:    s_mov_b32 s6, -1
; GFX12-FAKE16-NEXT:    s_mov_b32 s7, 0x31016000
; GFX12-FAKE16-NEXT:    s_mov_b32 s10, s6
; GFX12-FAKE16-NEXT:    s_mov_b32 s11, s7
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    s_mov_b32 s8, s2
; GFX12-FAKE16-NEXT:    s_mov_b32 s9, s3
; GFX12-FAKE16-NEXT:    s_mov_b32 s4, s0
; GFX12-FAKE16-NEXT:    buffer_load_b32 v0, off, s[8:11], null
; GFX12-FAKE16-NEXT:    s_mov_b32 s5, s1
; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x0
; GFX12-FAKE16-NEXT:    v_sqrt_bf16_e32 v1, v0
; GFX12-FAKE16-NEXT:    v_nop
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(TRANS32_DEP_2)
; GFX12-FAKE16-NEXT:    v_sqrt_bf16_e32 v0, v0
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
; GFX12-FAKE16-NEXT:    buffer_store_b32 v0, off, s[4:7], null
; GFX12-FAKE16-NEXT:    s_endpgm
entry:
  %a.val = load <2 x bfloat>, ptr addrspace(1) %a
  %r.val = call <2 x bfloat> @llvm.sqrt.v2bf16(<2 x bfloat> %a.val)
  store <2 x bfloat> %r.val, ptr addrspace(1) %r
  ret void
}